CARL Soliton Creation Demo

From The Artificial Experimentalist: Discovery and Control of Self-Organizing Phenomena with Autotelic Reinforcement Learning — Cvjetko, Hartl, Levin, Moulin-Frier & Oudeyer.

CARL is a goal-conditioned reinforcement learning agent that is able to guide the dynamics of Lenia (continuous cellular automaton) towards desired states through a sequence of local interventions. In this demo, CARL was trained to drive the grid's total mass towards some target value. At each step, CARL can select a spot on the board and either inject or erase mass there. The reward makes no reference to shape or structure, but because actions are costly, a self-sustaining configuration that matches the target is cheaper than intervening indefinitely. Soliton creation therefore emerges as a byproduct of reward maximization: the agent builds a localized, persistent pattern, a soliton, and stops acting once it holds the target mass on its own. Drag the target τ and watch CARL chase it, or pick a soliton recipe (hand-selected hyperparameters) to see what it can grow.

Hover the ? badges next to the controls for more detail.